Latency-Aware Soft-Voting Ensemble with Gradient Boosted Random Forest with Logistic Regression Hybrid in Clinical Decision Support System on Centralized EHR Infrastructure
Abstract
We present a latency-aware Clinical Decision Support System leveraging a softvoting
XGB, Random Forest, and Logistic Regression to improve performance and
operational efficiency in centralized Electronic Health Records environ. It explores
a de-identified dataset comprising of lab results, vital signs, prescriptions, and
demographic attributes – to evaluate predictive accuracy and end-to-end alert
latency. It exploits feature engineering and data normalization to ensure modelready
inputs that preserve clinical interpretability. The ensemble design fuses the
comprehensive knowledge of strong base learners, to capture of non-linear data
interactions, variance reduction, and probability calibration. Result shows the
ensemble outperforms benchmarks with 0.96 accuracy, 0.955 F1, 0.96 recall, and
AUC of 0.98. Latency evaluation demonstrated end-to-end processing times of
approximately 70 milliseconds, confirming the system’s suitability for near realtime
clinical applications. Comparative analysis highlights the ensemble’s
advantage over single learners, balancing sensitivity and specificity while
maintaining operational feasibility. These findings underscore the potential of
ensemble modes in enhancing CDSS reliability, robustness, and interpretability –
as crucial in centralized, real-time EHR infrastructures. The study contributes to
evidence that ensemble learning with optimized data preprocessing and
architecture-aware deployment can significantly improve predictive accuracy and
timely decision-making in clinical practice.